Claude Mythos #2: Cybersecurity and Project Glasswing

Anthropic will not release its most powerful new model, Claude Mythos, to the public for the time being, instead deploying it only to cybersecurity partners on a limited basis. The core purpose is to quickly find and patch vulnerabilities across major operating systems, browsers, and critical software in general through Project Glasswing.

Participating organizations will share their findings, and Anthropic has stated that this work is expected to continue for several months, with progress to be disclosed after 90 days. The deployment terms include $100 million worth of free credits and a $4 million cash donation, after which pricing will be set at around $25/$125 per million tokens.

The model card's cybersecurity description is very stark. It states that Mythos Preview can autonomously find zero-days in both open-source and closed-source software with minimal human involvement, and can develop a significant number of them into working proof-of-concept exploits.

Anthropic explains that because this capability has grown so significant, it is restricting access, and instead intends to have industry and open-source partners use it for defensive purposes. Had this been a general public release, strong detection and blocking measures would have been applied, but under the limited deployment, only relaxed monitoring is applied so as not to interfere with partners' actual security work.

In performance testing as well, most existing CTF benchmarks have largely become saturated, and CyberGym along with real-world code application are presented as exceptions and more realistic metrics. The author views these results not merely as a product launch, but as a turning point where the world's core software stack becomes dependent on AI's vulnerability analysis capabilities.

Governments had already been warned, but some senior officials complacently assumed this wouldn't happen, and the response was slow. As a result, industry moved first, and Mythos is being read as more than just a security tool—it is seen as a harbinger of automated AI R&D and even stronger future models.